(A) Schematic of visual word space, with one stored word (PENCIL) and two nonwords (PENICL and EPNCIL). We hypothesize that subjects would take longer to categorize a nonword when it is similar to a word, that is RT for PENICL would be larger than for EPNCIL. Thus, 1/RT would be proportional to this dissimilarity. Likewise we predicted that subjects would be faster to respond to frequent words which have a stronger stored representation. (B) Response times for words in the lexical decision task, sorted in descending order. The solid line represents the mean categorization time for words and the shaded bars represent s.e.m. Some example words are indicated using dotted lines. The split-half correlation between subjects (rsh) is indicated on the top. (C) Cross-validated model correlation between observed and predicted word response times across all words for various models: log word frequency (blue), number of orthographic neighbors (orange), log mean bigram frequency (purple), log mean letter frequency (cyan) and a combined model containing all these factors (red). Shaded error bars indicate mean ± sd of the correlation across 1000 random splits of the observed data. The asterisk indicates statistical significance of the comparison obtained by estimating the fraction of bootstrap samples in which the observed difference was violated (* is p<0.05, ** is p<0.005). (D) Response times for nonwords in the lexical decision task, sorted in descending order. Conventions as in (A). (E) Observed reciprocal response times for nonwords in the lexical decision task plotted against letter model predictions fit to the full data (450 nonwords). Some example nonwords are depicted. (F) Percent change in response time (nonword-RT – word-RT)/word-RT for middle and edge letter transpositions and for middle and edge substitutions for observed data (left) and for letter model predictions (right). MS: middle substitution. In both cases, asterisks represent statistical significance comparing the means of the corresponding groups using a rank-sum test (* is p<0.05, ** is p<0.005, etc.). (G) Observed reciprocal response times plotted against the Orthographic Levenshtein Distance (OLD), a popular model for edit distance between strings. (H) Cross-validated model correlation between observed and predicted nonword RTs for the letter model, OLD model, lexical model and the combined neural+lexical model. Conventions are as in (B).